Label identification result calibration method, device and equipment for communication infrastructure

Abstract

The invention discloses a label identification result calibration method, device and equipment of a communication infrastructure. The method comprises the following steps: acquiring an identification result of a label text of the communication infrastructure to be calibrated; performing address information separation processing on the identification result to obtain address information and non-address information; carrying out vectorization processing on the identification result to obtain a vector representation of the identification result; according to the recognition result, the address information, the non-address information and the vector representation of the recognition result, a stock facility database is retrieved to conduct comprehensive similarity matching, a similarity matching result of the recognition result is obtained, and the stock facility database comprises data tables of a plurality of types of communication infrastructures; the same data table comprises a plurality of label texts of the same type of communication infrastructure and corresponding vector representations; and obtaining a calibration result of the recognition result according to the similarity matching result. According to the invention, the calibration reliability can be improved, and the calibration effect is enhanced.

Description

Technical Field

[0001] This invention relates to the fields of communication and artificial intelligence technologies, and in particular to a method, apparatus, computer-readable storage medium, computer program product, and terminal equipment for calibrating tag identification results of communication infrastructure. Background Technology

[0002] The recognition effect of identifying label text from label images is limited by conditions such as lighting, viewing angle, pixels, and occlusions when acquiring the label images. Regardless of the recognition algorithm used, there will usually be a certain recognition error. Therefore, how to calibrate and correct the recognition results to make up for the recognition error and improve the accuracy and quality of the recognition results is a current research focus.

[0003] Manual calibration by comparing images is the most accurate method, but it is usually not feasible in actual production due to its high labor and time costs. Automated calibration has emerged as a solution. Among automated calibration methods, the most basic is rule-based calibration. This method performs well for recognition objects with simple and well-defined formats, such as license plate numbers, mobile phone numbers, ID card numbers, and device codes. It can directly extract the object to be recognized from the recognition results using formulas and regular expressions to distinguish the object from interference characters, or automatically repair missing and erroneous bits according to specific encoding rules. However, for more complex and diverse recognition objects, purely rule-based automated calibration methods have poor generalization, resulting in low reliability of calibration results. Summary of the Invention

[0004] The purpose of this invention is to provide a method, apparatus, computer-readable storage medium, computer program product, and terminal device for calibrating tag identification results of communication infrastructure. This method can solve the problem of low calibration reliability in existing automated calibration methods when facing complex and diverse identification objects, thereby improving the reliability of calibration and enhancing the calibration effect.

[0005] To achieve the above objectives, embodiments of the present invention provide a method for calibrating tag identification results of communication infrastructure, comprising:

[0006] Obtain the recognition results of the tag text of the communication infrastructure to be calibrated;

[0007] The identification result is subjected to address information separation processing to obtain address information and non-address information in the identification result;

[0008] The recognition result is vectorized to obtain a vector representation of the recognition result;

[0009] Based on the identification result, the address information, the non-address information, and the vector representation of the identification result, a preset database of existing facilities is retrieved to perform comprehensive similarity matching and obtain the similarity matching result of the identification result.

[0010] The calibration result of the recognition result is obtained based on the similarity matching result;

[0011] The existing infrastructure database includes data tables for several types of communication infrastructure, and each data table contains several tag texts and their corresponding vector representations for the same type of communication infrastructure.

[0012] Furthermore, the step of performing address information separation processing on the identification result to obtain address information and non-address information in the identification result specifically includes:

[0013] The recognition result is processed by a preset address extractor to separate address information and non-address information from the recognition result.

[0014] The address extractor is built based on the TinyBERT and CRF models and trained using an address recognition dataset constructed from address data of the region where the communication infrastructure is located.

[0015] Further, the step of vectorizing the recognition result to obtain a vector representation of the recognition result specifically includes:

[0016] The recognition results are vectorized according to a preset word vector model to obtain a vector representation of the recognition results;

[0017] The word vector model is built on the Word2Vec model and obtained through incremental training using labeled text from the communication infrastructure.

[0018] Furthermore, the existing facility database is constructed through the following steps:

[0019] Construct a blank database; wherein the blank database includes blank data tables for several types of communication infrastructure;

[0020] Extract the tag text for each type of communication infrastructure from the provincial resource system database;

[0021] The extracted label text of the same type of communication infrastructure and its vector representation obtained after vectorization are stored in the corresponding blank data table to obtain the data table for each type of communication infrastructure.

[0022] By aggregating the data tables of all types of communication infrastructure, the existing infrastructure database is obtained.

[0023] The types of communication infrastructure include computer rooms, racks, rack cabinets, network element equipment, and transmission optical paths. The tag text includes the name data of the communication infrastructure. The existing infrastructure database is updated according to a preset time period to synchronize the tag text in the provincial resource system database to the corresponding data table.

[0024] Furthermore, it also includes:

[0025] Obtain the facility type of the communication infrastructure to be calibrated;

[0026] Then, the step of retrieving a preset existing facility database based on the identification result, the address information, the non-address information, and the vector representation of the identification result to perform comprehensive similarity matching and obtain the similarity matching result of the identification result specifically includes:

[0027] In a pre-defined database of existing facilities, a data table of communication infrastructure of the same type as the facility type is identified and used as the target data table.

[0028] The recognition result, the address information, the non-address information, and the vector representation of the recognition result are compared with each tag text in the target data table and its corresponding vector representation to calculate the similarity, thereby obtaining the comprehensive similarity score between the recognition result and each tag text in the target data table.

[0029] The similarity matching result of the recognition result is obtained based on the comprehensive similarity score between the recognition result and all the tag texts in the target data table.

[0030] Further, the step of calculating the similarity between the recognition result, the address information, the non-address information, and the vector representation of the recognition result and each tag text in the target data table and its corresponding vector representation, to obtain a comprehensive similarity score between the recognition result and each tag text in the target data table, specifically includes:

[0031] The recognition result is compared with each tag text in the target data table to calculate the string similarity, and the first similarity score between the recognition result and each tag text in the target data table is obtained accordingly.

[0032] The address information is compared with each tag text in the target data table to calculate the string similarity, and the second similarity score between the recognition result and each tag text in the target data table is obtained accordingly.

[0033] The non-address information is compared with each tag text in the target data table to calculate the string similarity, and the third similarity score between the recognition result and each tag text in the target data table is obtained accordingly.

[0034] The vector representation of the recognition result is compared with the vector representation of each tag text in the target data table to calculate the cosine similarity, thereby obtaining the fourth similarity score between the recognition result and each tag text in the target data table.

[0035] Based on the first similarity score, second similarity score, third similarity score, and fourth similarity score corresponding to each tag text in the target data table, the comprehensive similarity score between the recognition result and each tag text in the target data table is obtained.

[0036] Further, the step of obtaining the comprehensive similarity score between the recognition result and each tag text in the target data table based on the first similarity score, second similarity score, third similarity score, and fourth similarity score corresponding to each tag text in the target data table specifically includes:

[0037] The overall similarity score between the recognition result and the i-th tag text in the target data table is calculated using the following formula:

[0038] Score(S,Pi)=w1*StrSim(S,Pi)+w2*StrSim(S1,Pi)+w3*StrSim(S2,Pi)+

[0039] w4*VecSim(norm(V),norm(Vi));

[0040] Where i represents the index of the tag text in the target data table, S represents the recognition result, S1 represents the address information, S2 represents the non-address information, V represents the vector representation of the recognition result, Pi represents the i-th tag text in the target data table, Vi represents the vector representation corresponding to Pi, norm() represents the normalization operation, StrSim(S,Pi) represents the first similarity score corresponding to Pi, StrSim(S1,Pi) represents the second similarity score corresponding to Pi, StrSim(S2,Pi) represents the third similarity score corresponding to Pi, VecSim(norm(V),norm(Vi)) represents the fourth similarity score corresponding to Pi, w1, w2, w3 and w4 are weight coefficients, and Score(S,Pi) represents the comprehensive similarity score between the recognition result and Pi.

[0041] Further, obtaining the calibration result of the recognition result based on the similarity matching result specifically includes:

[0042] Based on the similarity matching results, obtain the top-k results with the highest similarity and their corresponding comprehensive similarity scores; where k is a positive integer;

[0043] The comprehensive similarity scores corresponding to the top-k results are filtered according to a preset confidence threshold to obtain the calibration results of the recognition results.

[0044] To achieve the above objectives, embodiments of the present invention also provide a tag identification result calibration device for communication infrastructure, comprising:

[0045] The tag recognition result acquisition module is used to acquire the recognition results of the tag text of the communication infrastructure to be calibrated;

[0046] The address information separation module is used to perform address information separation processing on the identification result to obtain address information and non-address information in the identification result;

[0047] A vectorization processing module is used to perform vectorization processing on the recognition result to obtain a vector representation of the recognition result;

[0048] The label calibration and matching module is used to retrieve a preset existing facility database based on the identification result, the address information, the non-address information, and the vector representation of the identification result, so as to perform comprehensive similarity matching and obtain the similarity matching result of the identification result;

[0049] The label calibration result acquisition module is used to obtain the calibration result of the recognition result based on the similarity matching result;

[0050] The existing infrastructure database includes data tables for several types of communication infrastructure, and each data table contains several tag texts and their corresponding vector representations for the same type of communication infrastructure.

[0051] This invention also provides a computer-readable storage medium, including a stored computer program, which, when running, controls the device where the computer-readable storage medium is located to execute the tag identification result calibration method for the communication infrastructure described in any of the preceding embodiments.

[0052] This invention also provides a computer program product, including a computer program that, when executed by a processor, implements the tag identification result calibration method for any of the above-described communication infrastructures.

[0053] This invention also provides a terminal device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor, when executing the computer program, implements the tag identification result calibration method for the communication infrastructure described in any of the preceding claims.

[0054] Compared with existing technologies, embodiments of the present invention provide a method, apparatus, computer-readable storage medium, computer program product, and terminal device for calibrating tag identification results of communication infrastructure. First, the identification results of the tag text of the communication infrastructure to be calibrated are obtained. Next, address information separation processing is performed on the identification results to obtain address information and non-address information from the identification results, and the identification results are vectorized to obtain a vector representation of the identification results. Then, a preset existing infrastructure database is retrieved based on the identification results, address information, non-address information, and the vector representation of the identification results for comprehensive similarity matching to obtain a similarity matching result. The existing infrastructure database includes data tables for several types of communication infrastructure, and each data table includes several tag texts of the same type of communication infrastructure and their corresponding vector representations. Finally, the calibration result of the identification results is obtained based on the similarity matching result. Embodiments of the present invention can solve the problem of low calibration reliability in existing automated calibration methods when facing complex and diverse identification objects, thereby improving the reliability of calibration and enhancing the calibration effect. Attached Figure Description

[0055] Figure 1 This is a schematic diagram of a tag for a communication infrastructure provided in an embodiment of the present invention;

[0056] Figure 2 This is a flowchart of a tag identification result calibration method for communication infrastructure provided in an embodiment of the present invention;

[0057] Figure 3 This is a schematic diagram of a tag identification and calibration process for a communication infrastructure according to an embodiment of the present invention;

[0058] Figure 4 This is a structural block diagram of a tag identification result calibration device for communication infrastructure provided in an embodiment of the present invention;

[0059] Figure 5 This is a structural block diagram of a terminal device provided in an embodiment of the present invention. Detailed Implementation

[0060] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0061] It should be noted that in the intelligent application process for transmission network elements to enter the network, the initial stage requires collecting the tag text information from communication infrastructure such as equipment rooms, racks, racks, network element equipment, and transmission optical paths (transmission optical paths are physical channels in communication networks used to transmit optical signals, enabling high-speed data transmission between different network element devices through media such as optical fibers). This information is entered into the relevant system as an important part of the transmission network element's network access information. After the network access application is approved, the construction and installation steps are carried out. In this step, it is also necessary to collect the tag text information from communication infrastructure such as equipment rooms, racks, racks, network element equipment, and transmission optical paths to ensure that the resource information in the construction and installation stage is consistent with the resource information designed in the network access application. During actual network access operations, frontline operators take photos of tags in the equipment room, racks, rack cabinets, network element equipment, and transmission optical paths, and send them to the backend via a mobile app. The backend uses OCR (Optical Character Recognition) technology to recognize the tag text in the tag photos. However, due to limitations such as pixel count, lighting, viewing angle, obstructions, and other human error during photography, as well as the inherent accuracy of OCR technology, the tag text recognized from the tag photos may contain character errors, missing characters, or redundancy. Therefore, in order to improve the accuracy of network access information for transmission network elements and reduce secondary manual review, it is necessary to calibrate the tag recognition results of the communication infrastructure.

[0062] For example, see Figure 1 The diagram shown is a label schematic of a communication infrastructure provided in an embodiment of the present invention. Figure 1 Three types of tags on the communication infrastructure were shown (actually more than three types). Figure 1 The label shown in the upper left corner includes two parts: the carrier and the label text. Figure 1 The label shown in the lower left corner includes three parts: the carrier, the label text, and the QR code. Figure 1 The label shown on the right includes two parts: a QR code and label text. The operator's logo and name are displayed, while the label text indicates the name of the corresponding communication infrastructure. It typically consists of at least one character type, such as letters, numbers, or alphanumeric characters. For example... Figure 1The label text shown in the upper left corner reads "XX County XX Township Convergence Point 01F Computer Room 2 Columns 2 Racks A Side". Figure 1 The label text shown in the lower left corner reads "XX County XX College XX Campus Business Hub 01F Computer Room HN63423-1". Figure 1 The label text shown on the right is “102-0305-XX County XX Community-Gathering-SPN6700-24”. These label texts usually contain information such as the address, name, location, and number of the corresponding communication infrastructure.

[0063] It is evident that the format and content of the label text of these communication infrastructures are more complex and diverse than those of simple and clearly defined identification objects such as license plate numbers, mobile phone numbers, and ID card numbers. As described in the background section, the rule-based automated calibration method is not suitable for the business scenarios in this embodiment of the invention, and it is difficult to perform hard matching using specific encoding rules, fixed templates, enumeration, or other methods.

[0064] To address the aforementioned problems, embodiments of the present invention provide a method for calibrating tag identification results of communication infrastructure, see [link to relevant documentation]. Figure 2 The diagram shows a flowchart of a tag identification result calibration method for communication infrastructure according to an embodiment of the present invention. The method includes steps S11 to S15:

[0065] Step S11: Obtain the recognition results of the tag text of the communication infrastructure to be calibrated;

[0066] Step S12: Perform address information separation processing on the identification result to obtain address information and non-address information in the identification result;

[0067] Step S13: Perform vectorization processing on the recognition result to obtain the vector representation of the recognition result;

[0068] Step S14: Based on the identification result, the address information, the non-address information, and the vector representation of the identification result, retrieve the preset existing facility database to perform comprehensive similarity matching and obtain the similarity matching result of the identification result;

[0069] Step S15: Obtain the calibration result of the recognition result based on the similarity matching result;

[0070] The existing infrastructure database includes data tables for several types of communication infrastructure, and each data table contains several tag texts and their corresponding vector representations for the same type of communication infrastructure.

[0071] It should be noted that, in the embodiments of the present invention, an existing infrastructure database is constructed during service initialization. This existing infrastructure database includes data tables for several types of communication infrastructure. Each type of communication infrastructure corresponds to one data table, and the same data table includes several tag texts of the same type of communication infrastructure and the vector representation corresponding to each tag text (which can be obtained by vectorizing the tag text).

[0072] For example, communication infrastructure mainly includes five types: equipment rooms, racks, rack cabinets, network element equipment, and transmission optical paths. Therefore, the existing infrastructure database includes five data tables corresponding to these five types of communication infrastructure. Specifically, several tag texts corresponding to several equipment rooms, and the vector representation of each tag text, are stored in the data table corresponding to the equipment room; several tag texts corresponding to several racks, and the vector representation of each tag text, are stored in the data table corresponding to the rack; several tag texts corresponding to several rack cabinets, and the vector representation of each tag text, are stored in the data table corresponding to the rack; several tag texts corresponding to several network element equipment, and several tag texts corresponding to several transmission optical paths, and the vector representation of each tag text, are stored in the data table corresponding to the transmission optical path.

[0073] In specific implementation of this invention, firstly, the recognition result of the tag text of the communication infrastructure to be calibrated is obtained; then, address information separation processing is performed on the recognition result of the tag text of the communication infrastructure to be calibrated to extract the part describing the address from the recognition result of the tag text of the communication infrastructure to be calibrated, and correspondingly, address information and non-address information (non-address information is the remaining information other than address information) in the recognition result are obtained; then, the recognition result of the tag text of the communication infrastructure to be calibrated is vectorized to obtain the vector representation of the recognition result of the tag text of the communication infrastructure to be calibrated; then, based on the recognition result of the tag text of the communication infrastructure to be calibrated, the obtained address information, non-address information, and the vector representation of the recognition result, a pre-set existing facility database is retrieved, and using several tag texts stored in the existing facility database and the vector representation corresponding to each tag text, a comprehensive similarity matching is performed on the recognition result of the tag text of the communication infrastructure to be calibrated, and correspondingly, a similarity matching result of the recognition result of the tag text of the communication infrastructure to be calibrated is obtained; finally, the calibration result of the recognition result of the tag text of the communication infrastructure to be calibrated is obtained by further filtering from the obtained similarity matching result.

[0074] Specifically, obtaining the recognition results of the tag text of the communication infrastructure to be calibrated includes:

[0075] Acquire labeled images of the communication infrastructure to be calibrated;

[0076] The label text in the label image is recognized using OCR technology to obtain the recognition result.

[0077] Specifically, staff can first obtain images of the labels on the communication infrastructure to be calibrated by taking pictures, and then use OCR technology to recognize the label text in the captured image to obtain the recognition results of the label text of the communication infrastructure to be calibrated.

[0078] For example, the OCR technology used in the embodiments of the present invention can be implemented using an open-source framework (such as PaddleOCR). In order to enable the OCR model to better adapt to the data in this field, a certain amount of labeled images and corresponding labeled text of various communication infrastructures can be collected as annotation data to incrementally train the OCR model.

[0079] It should be noted that the calibration of the recognition result of the label text of the communication infrastructure as defined in the embodiments of the present invention is based on the fact that the recognition of the label text in the label has been completed (i.e., the recognition result of the label text of the communication infrastructure has been successfully obtained). Furthermore, in the actual recognition of the label text in the label, it is generally necessary to take a picture of the label first and then extract the label text from the label picture. This step can usually be implemented using OCR technology, and the OCR technology used is a mature existing technology, which will not be elaborated here.

[0080] It should be noted that the embodiments of the present invention are mainly applicable to the automated calibration of the recognition results of the label text of five types of communication infrastructure: computer rooms, racks, rack cabinets, network element equipment, and transmission optical paths. For other types of recognition objects, such as labels containing special formats such as documents with titles or tables, it is necessary not only to consider the calibration of the text part, but also to perform calibration of coordinates, orientation, and title level, so as to clarify the correspondence between each text and each cell and title. Since the label content of the communication infrastructure involved in the embodiments of the present invention does not contain these special formats, these other types of recognition objects are not within the scope of the embodiments of the present invention.

[0081] The present invention provides a calibration method for label recognition results of communication infrastructure. Targeting the specific characteristics of label text on communication infrastructure, it designs an address information separation mechanism, constructs a vectorized representation of the label text, and uses a database of existing infrastructure as the calibration basis. For the recognition results, address information, non-address information, and vector representation of the recognition results of the label text, a comprehensive matching mechanism integrating multiple similarity fusion methods is employed to automatically calibrate the recognition results of label text on communication infrastructure. This method solves the problem of low calibration reliability in existing automated calibration methods when facing complex and diverse recognition objects, thereby improving the reliability of calibration and achieving better calibration results.

[0082] It's worth noting that while using large image-text models (such as GLM) can achieve end-to-end recognition of tag text in tag images with high accuracy, the cost of deploying inference resources is relatively high, and the time overhead is difficult to meet business needs. Using external databases for calibration is a more reliable approach. For example, to identify the names of units and organizations in a specific province or city, a database of such names can be pre-collected and established. The recognition results of the tag text can then be compared with each entry in the database, and the highest similarity result (one or several entries) can be used as the calibration result. However, in various specific needs, factors such as the different recognition objects, the characteristics of the text to be recognized, the difference in recognition difficulty, the frequency of database updates, and the choice of similarity algorithm all constrain the calibration effect. Therefore, automated calibration methods relying on external databases need to be adapted and optimized for the specific characteristics of tags in communication infrastructure in terms of similarity matching. This mainly involves the following points:

[0083] (1) Due to technical limitations, external databases can only be updated at certain intervals and cannot be updated in real time. This may result in the tag information of the photo not being entered, thus mismatching other similar records in the database.

[0084] (2) Due to limitations in photographing conditions and the performance of the algorithm itself, OCR recognition results may be missing, redundant or incorrect, thus mismatching other similar records in the database.

[0085] (3) The unique semantic element combination and expression characteristics of the label text of communication infrastructure should be considered, rather than relying solely on word-level text features.

[0086] In response to the above situations, this embodiment of the invention improves the rationality of the calibration results through a combination of multiple means. In cases (1) and (2), it is generally desirable for the calibration procedure to give a lower similarity result in order to avoid the impact of erroneous high similarity on the accuracy of information entry. In these cases, relying solely on literal statistical features such as Levenshtein (edit distance), TF-IDF (Term Frequency-Inverse Document Frequency), and BM25 (Best Matching 25) is not reliable enough. This embodiment of the invention adopts a comprehensive matching mechanism that combines address information separation mechanism and multiple similarity fusion, so that when cases (1) and (2) occur, this embodiment of the invention can give a reasonable similarity result as much as possible. In case (3), this embodiment of the invention constructs a vectorized representation of the tag text, extracts the contextual semantic features of the tag text through a trainable model, and combines them with literal features, so that the calibration scheme can better adapt to the characteristics of the tag text, thereby improving the reliability of the calibration.

[0087] In one alternative embodiment, the existing facilities database is constructed through the following steps:

[0088] Construct a blank database; wherein the blank database includes blank data tables for several types of communication infrastructure;

[0089] Extract the tag text for each type of communication infrastructure from the provincial resource system database;

[0090] The extracted label text of the same type of communication infrastructure and its vector representation obtained after vectorization are stored in the corresponding blank data table to obtain the data table for each type of communication infrastructure.

[0091] By aggregating the data tables of all types of communication infrastructure, the existing infrastructure database is obtained.

[0092] The types of communication infrastructure include computer rooms, racks, rack cabinets, network element equipment, and transmission optical paths. The tag text includes the name data of the communication infrastructure. The existing infrastructure database is updated according to a preset time period to synchronize the tag text in the provincial resource system database to the corresponding data table.

[0093] It should be noted that since the basic resource data of communication infrastructure such as computer rooms, racks, rack cabinets, network element equipment and transmission optical paths have been collected in advance in the provincial resource system database and have been accumulated for a long time, the coverage and data accuracy can be relatively guaranteed. Therefore, the provincial resource system database can be used to construct the existing facility database used in the embodiments of the present invention.

[0094] Specifically, in conjunction with the above embodiments, the existing infrastructure database used in this embodiment of the invention can be constructed during service initialization through the following steps: First, a blank database is constructed, and blank data tables for several types of communication infrastructure are created in this blank database. For example, the types of communication infrastructure include computer rooms, racks, rack cabinets, network element equipment, and transmission optical paths. Then, five blank data tables corresponding to computer rooms, racks, rack cabinets, network element equipment, and transmission optical paths are created. Next, the tag text (which includes the name data of the communication infrastructure) for each type of communication infrastructure is extracted from the provincial resource system database, and the extracted tag text for the same type of communication infrastructure and the vector representation obtained after vectorizing the tag text are stored in the blank data table of the corresponding type of communication infrastructure, thus obtaining the data table for each type of communication infrastructure. Finally, the data tables for all types of communication infrastructure are collected together to obtain the existing infrastructure database.

[0095] It should be noted that when retrieving the existing infrastructure database based on the recognition results, address information, non-address information, and vector representations of the recognition results of the label text of the communication infrastructure to be calibrated for comprehensive similarity matching, vector matching is also involved. Therefore, it is necessary to convert each label text stored in the existing infrastructure database into a vector representation and store it. For example, the ElasticSearch open-source search engine can be used. Its 8.x version added support for vector search. When it is necessary to build / update the existing infrastructure database, the vector representation corresponding to each label text (which can be obtained by vectorizing the label text using a word vector model) can be inserted into the existing infrastructure database using the built-in functions of ElasticSearch.

[0096] It should be noted that, since the provincial resource system database is also continuously updated, once the business begins, the current existing facility database can be updated according to a pre-set time period (e.g., 00:00:00 every day) to periodically synchronize the tag text in the provincial resource system database to the corresponding data table.

[0097] It should be noted that, in combination Figure 1It is known that the labels of communication infrastructure may include other information besides the label text, such as operators and QR codes. However, these other information are not within the scope of the embodiments of the present invention. The embodiments of the present invention are preferably applicable to the automatic calibration of the recognition results of the label text (especially the device name in the label text). Therefore, when constructing the database of existing facilities, the embodiments of the present invention only need to extract the name data of five types of communication infrastructure, namely, computer room, rack, rack cabinet, network element equipment and transmission optical path, from the provincial resource system database.

[0098] In one optional embodiment, the step of performing address information separation processing on the identification result to obtain address information and non-address information in the identification result specifically includes:

[0099] The recognition result is processed by a preset address extractor to separate address information and non-address information from the recognition result.

[0100] The address extractor is built based on the TinyBERT and CRF models and trained using an address recognition dataset constructed from address data of the region where the communication infrastructure is located.

[0101] Specifically, in conjunction with the above embodiments, when performing address information separation processing on the recognition results of the label text of the communication infrastructure to be calibrated, a pre-set address extractor can be used to perform address information separation processing on the recognition results of the label text of the communication infrastructure to be calibrated, so as to extract the part describing the address from the recognition results of the label text of the communication infrastructure to be calibrated, and obtain the address information and non-address information in the recognition results accordingly.

[0102] It should be noted that the pre-set address extractor is the trained address extractor. Considering the deployment resources and inference latency requirements, you can choose to use the TinyBERT model and the CRF (conditional random field) model in series to build the address extractor. Then, collect address data of the region where the communication infrastructure is located (i.e. the region where the business is located) to build the address recognition dataset. After that, use the built address recognition dataset to train the built address extractor to obtain the trained address extractor.

[0103] For example, the address recognition dataset can be constructed by collecting the names of provinces, cities, districts, counties, and streets in the province where the business is located. Furthermore, for the smallest level addresses that cannot be exhaustively listed, such as "Building No. 3" or "Agricultural Bank of China XX Village Branch", a large amount of building, institution, and unit name data can be collected and added to the address recognition dataset.

[0104] In one optional embodiment, the step of vectorizing the recognition result to obtain a vector representation of the recognition result specifically includes:

[0105] The recognition results are vectorized according to a preset word vector model to obtain a vector representation of the recognition results;

[0106] The word vector model is built on the Word2Vec model and obtained through incremental training using labeled text from the communication infrastructure.

[0107] Specifically, in conjunction with the above embodiments, when vectorizing the recognition results of the label text of the communication infrastructure to be calibrated, a preset word vector model can be used to vectorize the recognition results of the label text of the communication infrastructure to be calibrated, thereby obtaining a vector representation of the recognition results of the label text of the communication infrastructure to be calibrated.

[0108] It should be noted that the pre-set word vector model is the trained word vector model. Considering the characteristics of the labeled text of communication infrastructure, the word vector model can be built based on the Word2Vec model, and incremental training can be performed using a large amount of labeled text of communication infrastructure to obtain the trained word vector model.

[0109] For example, the open-source lightweight Chinese word vector model w2v-light-tencent-chinese can be used to build the word vector model. This model is based on Word2Vec and can convert words into fixed 1×100-dimensional vector representations. The depth of feature extraction is between pure literal statistical features and pure semantic features. Compared with edit distance, bag-of-words model, etc., this model can better model the contextual dependencies in word sequences and discover the implicit associations between fields in the tagged text. Compared with sentence vector models that focus on complete semantics, this model focuses more on capturing word-level features, and is therefore suitable for the business scenarios of this embodiment of the invention.

[0110] In one alternative embodiment, the method further includes:

[0111] Obtain the facility type of the communication infrastructure to be calibrated;

[0112] Then, the step of retrieving a preset existing facility database based on the identification result, the address information, the non-address information, and the vector representation of the identification result to perform comprehensive similarity matching and obtain the similarity matching result of the identification result specifically includes:

[0113] In a pre-defined database of existing facilities, a data table of communication infrastructure of the same type as the facility type is identified and used as the target data table.

[0114] The recognition result, the address information, the non-address information, and the vector representation of the recognition result are compared with each tag text in the target data table and its corresponding vector representation to calculate the similarity, thereby obtaining the comprehensive similarity score between the recognition result and each tag text in the target data table.

[0115] The similarity matching result of the recognition result is obtained based on the comprehensive similarity score between the recognition result and all the tag texts in the target data table.

[0116] Specifically, in conjunction with the above embodiments, when searching the pre-set existing infrastructure database, since the existing infrastructure database stores data tables of different types of communication infrastructure, and the data tables of different types of communication infrastructure store tag texts and their corresponding vector representations of different types of communication infrastructure, if the identification result, address information, non-address information, and vector representation of the identification result of the communication infrastructure to be calibrated are sequentially matched with each tag text and its corresponding vector representation stored in the existing infrastructure database for similarity, it will result in a large amount of data processing and affect the matching efficiency. Therefore, the search in the existing infrastructure database can be targeted based on the facility type of the communication infrastructure to be calibrated. That is, only the data tables of communication infrastructures in the existing infrastructure database that are the same as the facility type of the communication infrastructure to be calibrated need to be searched, and there is no need to search the data tables of communication infrastructures in the existing infrastructure database that are different from the facility type of the communication infrastructure to be calibrated.

[0117] Furthermore, in specific implementation, the facility type of the communication infrastructure to be calibrated can be obtained from the business layer first. Then, based on the facility type of the communication infrastructure to be calibrated, a data table of communication infrastructures with the same type as the facility type of the communication infrastructure to be calibrated can be found in the existing facility database, and the found data table can be used as the target data table. Next, the identification result, address information, non-address information, and vector representation of the identification result of the communication infrastructure to be calibrated are compared with each tag text and its corresponding vector representation stored in the target data table to calculate the similarity. The comprehensive similarity score between the identification result of the communication infrastructure to be calibrated and each tag text in the target data table can be obtained. Finally, the comprehensive similarity scores of the identification result of the communication infrastructure to be calibrated and all the tag texts in the target data table are combined to obtain the similarity matching result of the identification result of the communication infrastructure to be calibrated.

[0118] It should be noted that in the early stage of label text recognition (e.g., OCR recognition stage), the work order process will be designed to recognize the label text of a certain type of communication infrastructure in sequence. Therefore, when uploading the label image of a certain type of communication infrastructure, the facility type of the corresponding communication infrastructure will also be sent to the OCR recognition interface. After the OCR recognition is completed, the business layer will send the facility type of the communication infrastructure and the recognition result to the next processing stage (i.e., the label recognition result calibration stage as defined in this embodiment of the invention).

[0119] In one optional embodiment, the step of calculating the similarity between the recognition result, the address information, the non-address information, and the vector representation of the recognition result and each tag text in the target data table and its corresponding vector representation, respectively, to obtain a comprehensive similarity score between the recognition result and each tag text in the target data table, specifically includes:

[0120] The recognition result is compared with each tag text in the target data table to calculate the string similarity, and the first similarity score between the recognition result and each tag text in the target data table is obtained accordingly.

[0121] The address information is compared with each tag text in the target data table to calculate the string similarity, and the second similarity score between the recognition result and each tag text in the target data table is obtained accordingly.

[0122] The non-address information is compared with each tag text in the target data table to calculate the string similarity, and the third similarity score between the recognition result and each tag text in the target data table is obtained accordingly.

[0123] The vector representation of the recognition result is compared with the vector representation of each tag text in the target data table to calculate the cosine similarity, thereby obtaining the fourth similarity score between the recognition result and each tag text in the target data table.

[0124] Based on the first similarity score, second similarity score, third similarity score, and fourth similarity score corresponding to each tag text in the target data table, the comprehensive similarity score between the recognition result and each tag text in the target data table is obtained.

[0125] Specifically, in conjunction with the above embodiments, when obtaining the comprehensive similarity score between the identification result of the communication infrastructure to be calibrated and each tag text in the target data table, the identification result of the communication infrastructure to be calibrated can first be compared with each tag text in the target data table to calculate the string similarity, thereby obtaining the first similarity score between the identification result of the communication infrastructure to be calibrated and each tag text in the target data table; similarly, the address information in the identification result of the communication infrastructure to be calibrated can be compared with each tag text in the target data table to calculate the string similarity, thereby obtaining the second similarity score between the identification result of the communication infrastructure to be calibrated and each tag text in the target data table; the non-address information in the identification result of the communication infrastructure to be calibrated can be compared with the target data table to calculate the string similarity, thereby obtaining the second similarity score between the identification result of the communication infrastructure to be calibrated and each tag text in the target data table; String similarity is calculated for each tag text in the table to obtain the third similarity score between the identification result of the communication infrastructure to be calibrated and each tag text in the target data table. Cosine similarity is calculated between the vector representation of the identification result of the communication infrastructure to be calibrated and the vector representation of each tag text in the target data table to obtain the fourth similarity score between the identification result of the communication infrastructure to be calibrated and each tag text in the target data table. Then, the first, second, third, and fourth similarity scores corresponding to each tag text in the target data table are weighted and summed to obtain the comprehensive similarity score between the identification result of the communication infrastructure to be calibrated and each tag text in the target data table.

[0126] In one optional embodiment, obtaining the comprehensive similarity score between the recognition result and each tag text in the target data table based on the first similarity score, second similarity score, third similarity score, and fourth similarity score corresponding to each tag text in the target data table specifically includes:

[0127] The overall similarity score between the recognition result and the i-th tag text in the target data table is calculated using the following formula:

[0128] Score(S,Pi)=w1*StrSim(S,Pi)+w2*StrSim(S1,Pi)+w3*StrSim(S2,Pi)+

[0129] w4*VecSim(norm(V),norm(Vi));

[0130] Where i represents the index of the tag text in the target data table, S represents the recognition result, S1 represents the address information, S2 represents the non-address information, V represents the vector representation of the recognition result, Pi represents the i-th tag text in the target data table, Vi represents the vector representation corresponding to Pi, norm() represents the normalization operation, StrSim(S,Pi) represents the first similarity score corresponding to Pi, StrSim(S1,Pi) represents the second similarity score corresponding to Pi, StrSim(S2,Pi) represents the third similarity score corresponding to Pi, VecSim(norm(V),norm(Vi)) represents the fourth similarity score corresponding to Pi, w1, w2, w3 and w4 are weight coefficients, and Score(S,Pi) represents the comprehensive similarity score between the recognition result and Pi.

[0131] Specifically, in conjunction with the above embodiments, the principle for obtaining the comprehensive similarity score between the identification result of the communication infrastructure to be calibrated and each tag text in the target data table is the same. Here, we take the i-th tag text in the target data table as an example. That is, for the i-th tag text in the target data table, the comprehensive similarity score between the identification result of the communication infrastructure to be calibrated and the i-th tag text in the target data table can be directly calculated using the following formula:

[0132] Score(S,Pi)=w1*StrSim(S,Pi)+w2*StrSim(S1,Pi)+w3*StrSim(S2,Pi)+

[0133] w4*VecSim(norm(V),norm(Vi));

[0134] Where S represents the identification result of the communication infrastructure to be calibrated, S1 represents the address information in the identification result of the communication infrastructure to be calibrated, S2 represents the non-address information in the identification result of the communication infrastructure to be calibrated, V represents the vector representation of the identification result of the communication infrastructure to be calibrated, Pi represents the i-th tag text in the target data table, Vi represents the vector representation of Pi, norm() represents the normalization operation, StrSim(S,Pi) represents the first similarity score of Pi, StrSim(S1,Pi) represents the second similarity score of Pi, StrSim(S2,Pi) represents the third similarity score of Pi, VecSim(norm(V),norm(Vi)) represents the fourth similarity score of Pi, w1, w2, w3 and w4 are weight coefficients, and Score(S,Pi) represents the comprehensive similarity score between the identification result of the communication infrastructure to be calibrated and Pi.

[0135] It should be noted that StrSim() can use the fuzzy query method built into ElasticSearch and normalize the similarity score to the [0, 1] interval for summing with VecSim(); VecSim() can use cosine similarity; norm() can use L2 norm normalization of the vector representation, and since V and Vi have been normalized, the value range of VecSim() is also in the [0, 1] interval, which is convenient for summing with StrSim(); w1, w2, w3 and w4 are the weight coefficients of the first similarity score, the second similarity score, the third similarity score and the fourth similarity score, respectively. The values ​​of each weight coefficient can be adjusted according to the requirements and the actual test results. This embodiment of the invention does not impose specific limitations.

[0136] In one optional embodiment, obtaining the calibration result of the recognition result based on the similarity matching result specifically includes:

[0137] Based on the similarity matching results, obtain the top-k results with the highest similarity and their corresponding comprehensive similarity scores; where k is a positive integer;

[0138] The comprehensive similarity scores corresponding to the top-k results are filtered according to a preset confidence threshold to obtain the calibration results of the recognition results.

[0139] Specifically, in conjunction with the above embodiments, when further filtering the obtained similarity matching results to obtain the calibration results of the recognition results of the label text of the communication infrastructure to be calibrated, the top-k results with the highest similarity (the top-k results are the top k and bottom k results of all comprehensive similarity scores in the similarity matching results, sorted from largest to smallest / smallest to largest) and the comprehensive similarity scores corresponding to the top-k results can be obtained again according to the pre-set confidence threshold. Then, the comprehensive similarity scores corresponding to the top-k results can be filtered again to obtain the comprehensive similarity scores (there may be more than one) that exceed the confidence threshold. The label texts corresponding to these comprehensive similarity scores that exceed the confidence threshold in the target data table can be further determined. The label texts further determined in the target data table can then be used as the calibration results of the recognition results of the label text of the communication infrastructure to be calibrated.

[0140] It should be noted that k is a positive integer, which can be k=1 or k>1, and the specific value of k can be adjusted according to business needs. In order to simplify the calibration results, k=1 can be used, that is, the highest comprehensive similarity score in the similarity matching results can be directly selected and the confidence score can be further filtered.

[0141] It should be noted that the specific value of the confidence threshold can also be adjusted according to business needs. By setting the confidence threshold to filter the top-k results, the reliability of the calibration results can be further improved.

[0142] For example, see Figure 3 The diagram shown is a schematic representation of a tag identification and calibration process for a communication infrastructure according to an embodiment of the present invention. Figure 3 The portion within the dashed box shown represents the specific implementation process of this invention. The overall process of tag identification and calibration is as follows: After the process begins, staff take a photo of a tag on the communication infrastructure and upload it to the business layer. The business layer first calls the tag identification OCR capability to perform OCR recognition on the tag photo and outputs the OCR recognition result of the tag photo accordingly. Then, using the automated calibration scheme provided in this invention, the OCR recognition result of the tag photo is subjected to comprehensive similarity matching. The comprehensive similarity matching process requires searching the same type of data table in the existing facility database (which is periodically synchronized with the provincial resource system database) based on the facility type of the communication infrastructure and the OCR recognition result, and outputting the similarity matching result of the OCR recognition result accordingly. After that, the similarity matching result of the OCR recognition result is filtered (i.e., top-k result filtering and confidence filtering are performed), and the calibration result is output and returned accordingly. At this point, the entire process ends.

[0143] It should be noted that the automated calibration service provided in this embodiment of the invention exposes two functional interfaces: Interface 1: used when establishing / updating the existing facility database; Interface 2: used to receive tag photos and return calibration results.

[0144] It should be noted that the business scenarios involved in the embodiments of the present invention have some characteristics, which, compared with more general scenarios, bring the following challenges to the reliability of calibration:

[0145] On the one hand, in transmission network services, equipment and facilities such as equipment rooms, cabinets, racks, network element equipment, and transmission optical paths are constantly being added. However, from a technical implementation perspective, the calibration scheme cannot receive notifications every time a new data entry occurs in the provincial resource system database and pull data to the existing facility database in real time. It can only periodically and proactively obtain incremental data from the provincial side. During the interval between two updates, there may be situations where the tag information to be identified has not yet been entered. In this case, it may match unexpected records in the existing facility database. For example, Table 1 gives some examples of errors in tag identification and calibration. As shown in Example 1, although the OCR correctly identified the actual content of the tag as "A County A1 Township Convergence Point 01F Equipment Room 2 Column 2 Rack A Side", since this record has not yet been entered into the database, it matched the record with the highest similarity in the database as "B City B1 Community Convergence Point 01F Equipment Room / 2 Column - 2 Rack A Side", which is an incorrect calibration result. Example 2 also has a similar problem, where the calibration scheme incorrectly matches a record with the same address as the target but on a different rack. For Examples 1 and 2, the OCR recognition results and erroneous calibration results have few repeated characters and low text similarity, which can alert staff that there may be a calibration deviation. However, in Examples 3 and 4, because there are records in the existing facility database that are very similar to the actual labels, but have fundamental differences in key locations such as addresses and rack numbers, these records are incorrectly matched and output with high similarity, which is inconsistent with the calibration expectations.

[0146] Table 1

[0147]

[0148]

[0149] On the other hand, even if the existing facility database contains the correct label records, missing, redundant, or erroneous OCR recognition results can still lead to incorrect calibration results. Examples 5, 6, and 7 in Table 1 illustrate cases of missing, redundant, and erroneous OCR results, respectively. Although the causes differ, they cause similar problems to Examples 3 and 4, all affecting the reliability of the calibration results.

[0150] Furthermore, the label text for communication infrastructure follows certain patterns and paradigms. For example, the address is usually at the beginning and arranged in descending order, while the specific facility name, location, or number of the rack, optical path, or network element is at the end. In the label text for racks, the name of the equipment room where the rack is located is usually listed first. In the label text for rack units, the order is generally: equipment room, system, rack voltage, and rack number, but some items may be missing. In addition, different fields in the label text are interrelated. For example, the DC rack voltage and the types of transmission network elements used in large data center equipment rooms and aggregation / access layer equipment rooms are usually different. Therefore, these non-explicit association patterns in the label text also need to be considered.

[0151] Based on the above analysis of the characteristics of tag recognition scenarios for communication infrastructure and the challenges they pose to calibration results, this invention provides a calibration method for tag recognition results of communication infrastructure. Although it adopts a relatively common calibration scheme based on external databases and text similarity matching, this invention has designed specific optimizations and improvements tailored to the specific characteristics of tag text in communication infrastructure. These improvements mainly include an address information separation mechanism, the construction of a vectorized representation of tag text, and a comprehensive matching mechanism that integrates multiple similarities. These improvements enhance the reliability of calibration and achieve better calibration results.

[0152] Furthermore, this embodiment of the invention is geared towards the network access scenario of transmission network elements. It utilizes the provincial-level resource system database to construct an existing facility database to calibrate the OCR recognition results and regularly perform incremental updates to the existing facility database. This can compensate for the limitations of the OCR algorithm itself in terms of recognition effect, ensure the accuracy of the information entered into the network, save the cost of manual review and verification of information entry, and effectively serve actual production, thereby improving the efficiency of resource network access.

[0153] It should be noted that the embodiments of the present invention also describe a closed-loop process including tag photo acquisition and tag text recognition, calibration database establishment and regular updates, text similarity calculation, and calibration result output, as well as open-source tools that can be referenced for each step, which are feasible for implementation and deployment. Furthermore, the embodiments of the present invention can be built on an open-source framework, have low requirements for hardware resources, are easy to deploy and use, and have a certain degree of wide applicability.

[0154] This invention also provides a tag identification result calibration device for communication infrastructure, used to implement the tag identification result calibration method for communication infrastructure described in any of the above embodiments. See [link to relevant documentation]. Figure 4 The diagram shown is a structural block diagram of a tag identification result calibration device for communication infrastructure according to an embodiment of the present invention. The device includes:

[0155] The tag recognition result acquisition module 11 is used to acquire the recognition result of the tag text of the communication infrastructure to be calibrated;

[0156] Address information separation module 12 is used to perform address information separation processing on the identification result to obtain address information and non-address information in the identification result;

[0157] The vectorization processing module 13 is used to perform vectorization processing on the recognition result to obtain a vector representation of the recognition result;

[0158] The tag calibration and matching module 14 is used to retrieve a preset existing facility database based on the identification result, the address information, the non-address information and the vector representation of the identification result, so as to perform comprehensive similarity matching and obtain the similarity matching result of the identification result;

[0159] The label calibration result acquisition module 15 is used to acquire the calibration result of the recognition result based on the similarity matching result;

[0160] The existing infrastructure database includes data tables for several types of communication infrastructure, and each data table contains several tag texts and their corresponding vector representations for the same type of communication infrastructure.

[0161] Preferably, the address information separation module 12 is specifically used for:

[0162] The recognition result is processed by a preset address extractor to separate address information and non-address information from the recognition result.

[0163] The address extractor is built based on the TinyBERT and CRF models and trained using an address recognition dataset constructed from address data of the region where the communication infrastructure is located.

[0164] Preferably, the vectorization processing module 13 is specifically used for:

[0165] The recognition results are vectorized according to a preset word vector model to obtain a vector representation of the recognition results;

[0166] The word vector model is built on the Word2Vec model and obtained through incremental training using labeled text from the communication infrastructure.

[0167] Preferably, the apparatus further includes a database construction module, which is used to construct the existing facility database through the following steps:

[0168] Construct a blank database; wherein the blank database includes blank data tables for several types of communication infrastructure;

[0169] Extract the tag text for each type of communication infrastructure from the provincial resource system database;

[0170] The extracted label text of the same type of communication infrastructure and its vector representation obtained after vectorization are stored in the corresponding blank data table to obtain the data table for each type of communication infrastructure.

[0171] By aggregating the data tables of all types of communication infrastructure, the existing infrastructure database is obtained.

[0172] The types of communication infrastructure include computer rooms, racks, rack cabinets, network element equipment, and transmission optical paths. The tag text includes the name data of the communication infrastructure. The existing infrastructure database is updated according to a preset time period to synchronize the tag text in the provincial resource system database to the corresponding data table.

[0173] Preferably, the device further includes:

[0174] Facility type acquisition module, used to acquire the facility type of the communication infrastructure to be calibrated;

[0175] Therefore, the tag calibration and matching module 14 specifically includes:

[0176] The matching database determination unit is used to determine, in a preset existing facility database, a data table of communication infrastructure of the same type as the facility type, and use it as the target data table;

[0177] The label calibration and matching unit is used to calculate the similarity between the recognition result, the address information, the non-address information and the vector representation of the recognition result and each label text and its corresponding vector representation in the target data table, respectively, and obtain the comprehensive similarity score between the recognition result and each label text in the target data table.

[0178] The matching result acquisition unit is used to obtain the similarity matching result of the recognition result based on the comprehensive similarity score between the recognition result and all the tag texts in the target data table.

[0179] Preferably, the tag calibration matching unit specifically includes:

[0180] The first similarity calculation subunit is used to calculate the string similarity between the recognition result and each tag text in the target data table, and obtain the first similarity score between the recognition result and each tag text in the target data table.

[0181] The second similarity calculation subunit is used to perform string similarity calculation between the address information and each tag text in the target data table, and obtain the second similarity score between the recognition result and each tag text in the target data table.

[0182] The third similarity calculation subunit is used to calculate the string similarity between the non-address information and each tag text in the target data table, and obtain the third similarity score between the recognition result and each tag text in the target data table.

[0183] The fourth similarity calculation subunit is used to perform cosine similarity calculation between the vector representation of the recognition result and the vector representation corresponding to each tag text in the target data table, and obtain the fourth similarity score between the recognition result and each tag text in the target data table.

[0184] The comprehensive similarity calculation subunit is used to obtain the comprehensive similarity score between the recognition result and each tag text in the target data table based on the first similarity score, second similarity score, third similarity score and fourth similarity score corresponding to each tag text in the target data table.

[0185] Preferably, the comprehensive similarity calculation subunit is specifically used for:

[0186] The overall similarity score between the recognition result and the i-th tag text in the target data table is calculated using the following formula:

[0187] Score(S,Pi)=w1*StrSim(S,Pi)+w2*StrSim(S1,Pi)+w3*StrSim(S2,Pi)+

[0188] w4*VecSim(norm(V),norm(Vi));

[0189] Where i represents the index of the tag text in the target data table, S represents the recognition result, S1 represents the address information, S2 represents the non-address information, V represents the vector representation of the recognition result, Pi represents the i-th tag text in the target data table, Vi represents the vector representation corresponding to Pi, norm() represents the normalization operation, StrSim(S,Pi) represents the first similarity score corresponding to Pi, StrSim(S1,Pi) represents the second similarity score corresponding to Pi, StrSim(S2,Pi) represents the third similarity score corresponding to Pi, VecSim(norm(V),norm(Vi)) represents the fourth similarity score corresponding to Pi, w1, w2, w3 and w4 are weight coefficients, and Score(S,Pi) represents the comprehensive similarity score between the recognition result and Pi.

[0190] Preferably, the label calibration result acquisition module 15 specifically includes:

[0191] The first calibration result acquisition unit is used to acquire the top-k results with the highest similarity and their corresponding comprehensive similarity scores based on the similarity matching results; where k is a positive integer;

[0192] The second calibration result acquisition unit is used to filter the comprehensive similarity scores corresponding to the top-k results according to a preset confidence threshold, and obtain the calibration result of the recognition result.

[0193] It should be noted that the tag identification result calibration device for communication infrastructure provided in this embodiment of the invention can realize all the processes of the tag identification result calibration method for communication infrastructure described in any of the above embodiments. The functions and technical effects of each module, unit, and subunit in the device are the same as the functions and technical effects of the tag identification result calibration method for communication infrastructure described in the above embodiments, and will not be repeated here.

[0194] This invention also provides a computer-readable storage medium including a stored computer program, which, when running, controls the device where the computer-readable storage medium is located to execute the tag identification result calibration method for the communication infrastructure described in any of the above embodiments.

[0195] This invention also provides a computer program product, including a computer program that, when executed by a processor, implements the tag identification result calibration method for communication infrastructure described in any of the above embodiments.

[0196] This invention also provides a terminal device, see [link to relevant documentation]. Figure 5The diagram shown is a structural block diagram of a terminal device according to an embodiment of the present invention. The terminal device includes a processor 10, a memory 20, and a computer program stored in the memory 20 and configured to be executed by the processor 10. When the processor 10 executes the computer program, it implements the tag identification result calibration method for the communication infrastructure described in any of the above embodiments.

[0197] Preferably, the computer program can be divided into one or more modules / units (such as computer program 1, computer program 2, ...), and the one or more modules / units are stored in the memory 20 and executed by the processor 10 to complete the present invention. The one or more modules / units can be a series of computer program instruction segments capable of performing specific functions, and the instruction segments are used to describe the execution process of the computer program in the terminal device.

[0198] The processor 10 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor, or the processor 10 may be any conventional processor. The processor 10 is the control center of the terminal device, connecting various parts of the terminal device through various interfaces and lines.

[0199] The memory 20 mainly includes a program storage area and a data storage area. The program storage area can store the operating system, applications required for at least one function, etc., while the data storage area can store related data, etc. Furthermore, the memory 20 can be a high-speed random access memory, or a non-volatile memory, such as a plug-in hard disk, a smart media card (SMC), a secure digital card (SD), and a flash card, or other volatile solid-state storage devices.

[0200] It should be noted that the aforementioned terminal devices may include, but are not limited to, processors and memory, as will be understood by those skilled in the art. Figure 5The structural block diagram shown is merely a structural example of the terminal device described above and does not constitute a limitation on the structure of the terminal device. The terminal device may include more or fewer components than shown, or combine certain components, or use different components.

[0201] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for calibrating tag identification results of communication infrastructure, characterized in that, include: Obtain the recognition results of the tag text of the communication infrastructure to be calibrated; The identification result is subjected to address information separation processing to obtain address information and non-address information in the identification result; The recognition result is vectorized to obtain a vector representation of the recognition result; Based on the identification result, the address information, the non-address information, and the vector representation of the identification result, a preset database of existing facilities is retrieved to perform comprehensive similarity matching and obtain the similarity matching result of the identification result. The calibration result of the recognition result is obtained based on the similarity matching result; The existing infrastructure database includes data tables for several types of communication infrastructure, and each data table contains several tag texts and their corresponding vector representations for the same type of communication infrastructure.

2. The tag identification result calibration method for communication infrastructure as described in claim 1, characterized in that, The step of performing address information separation processing on the identification result to obtain address information and non-address information in the identification result specifically includes: The recognition result is processed by a preset address extractor to separate address information and non-address information from the recognition result. The address extractor is built based on the TinyBERT and CRF models and trained using an address recognition dataset constructed from address data of the region where the communication infrastructure is located.

3. The tag identification result calibration method for communication infrastructure as described in claim 1, characterized in that, The step of vectorizing the recognition result to obtain a vector representation of the recognition result specifically includes: The recognition results are vectorized according to a preset word vector model to obtain a vector representation of the recognition results; The word vector model is built on the Word2Vec model and obtained through incremental training using labeled text from the communication infrastructure.

4. The tag identification result calibration method for communication infrastructure as described in claim 1, characterized in that, The existing infrastructure database is constructed through the following steps: Construct a blank database; wherein the blank database includes blank data tables for several types of communication infrastructure; Extract the tag text for each type of communication infrastructure from the provincial resource system database; The extracted label text of the same type of communication infrastructure and its vector representation obtained after vectorization are stored in the corresponding blank data table to obtain the data table for each type of communication infrastructure. By aggregating the data tables of all types of communication infrastructure, the existing infrastructure database is obtained. The types of communication infrastructure include computer rooms, racks, rack cabinets, network element equipment, and transmission optical paths. The tag text includes the name data of the communication infrastructure. The existing infrastructure database is updated according to a preset time period to synchronize the tag text in the provincial resource system database to the corresponding data table.

5. The tag identification result calibration method for communication infrastructure as described in claim 1, characterized in that, Also includes: Obtain the facility type of the communication infrastructure to be calibrated; Then, the step of retrieving a preset existing facility database based on the identification result, the address information, the non-address information, and the vector representation of the identification result to perform comprehensive similarity matching and obtain the similarity matching result of the identification result specifically includes: In a pre-defined database of existing facilities, a data table of communication infrastructure of the same type as the facility type is identified and used as the target data table. The recognition result, the address information, the non-address information, and the vector representation of the recognition result are compared with each tag text in the target data table and its corresponding vector representation to calculate the similarity, thereby obtaining the comprehensive similarity score between the recognition result and each tag text in the target data table. The similarity matching result of the recognition result is obtained based on the comprehensive similarity score between the recognition result and all the tag texts in the target data table.

6. The tag identification result calibration method for communication infrastructure as described in claim 5, characterized in that, The step of calculating the similarity between the recognition result, the address information, the non-address information, and the vector representation of the recognition result, and each tag text in the target data table and its corresponding vector representation, to obtain a comprehensive similarity score between the recognition result and each tag text in the target data table, specifically includes: The recognition result is compared with each tag text in the target data table to calculate the string similarity, and the first similarity score between the recognition result and each tag text in the target data table is obtained accordingly. The address information is compared with each tag text in the target data table to calculate the string similarity, and the second similarity score between the recognition result and each tag text in the target data table is obtained accordingly. The non-address information is compared with each tag text in the target data table to calculate the string similarity, and the third similarity score between the recognition result and each tag text in the target data table is obtained accordingly. The vector representation of the recognition result is compared with the vector representation of each tag text in the target data table to calculate the cosine similarity, thereby obtaining the fourth similarity score between the recognition result and each tag text in the target data table. Based on the first similarity score, second similarity score, third similarity score, and fourth similarity score corresponding to each tag text in the target data table, the comprehensive similarity score between the recognition result and each tag text in the target data table is obtained.

7. The tag identification result calibration method for communication infrastructure as described in claim 6, characterized in that, The step of obtaining a comprehensive similarity score between the recognition result and each tag text in the target data table based on the first similarity score, second similarity score, third similarity score, and fourth similarity score corresponding to each tag text in the target data table specifically includes: The overall similarity score between the recognition result and the i-th tag text in the target data table is calculated using the following formula: Score(S,Pi)=w1*StrSim(S,Pi)+w2*StrSim(S1,Pi)+w3*StrSim(S2,Pi)+ w4*VecSim(norm(V),norm(Vi)); Where i represents the index of the tag text in the target data table, S represents the recognition result, S1 represents the address information, S2 represents the non-address information, V represents the vector representation of the recognition result, Pi represents the i-th tag text in the target data table, Vi represents the vector representation corresponding to Pi, norm() represents the normalization operation, StrSim(S,Pi) represents the first similarity score corresponding to Pi, StrSim(S1,Pi) represents the second similarity score corresponding to Pi, StrSim(S2,Pi) represents the third similarity score corresponding to Pi, VecSim(norm(V),norm(Vi)) represents the fourth similarity score corresponding to Pi, w1, w2, w3 and w4 are weight coefficients, and Score(S,Pi) represents the comprehensive similarity score between the recognition result and Pi.

8. The tag identification result calibration method for communication infrastructure as described in claim 5, characterized in that, The step of obtaining the calibration result of the recognition result based on the similarity matching result specifically includes: Based on the similarity matching results, obtain the top-k results with the highest similarity and their corresponding comprehensive similarity scores; where k is a positive integer; The comprehensive similarity scores corresponding to the top-k results are filtered according to a preset confidence threshold to obtain the calibration results of the recognition results.

9. A tag identification result calibration device for communication infrastructure, characterized in that, include: The tag recognition result acquisition module is used to acquire the recognition results of the tag text of the communication infrastructure to be calibrated; The address information separation module is used to perform address information separation processing on the identification result to obtain address information and non-address information in the identification result; A vectorization processing module is used to perform vectorization processing on the recognition result to obtain a vector representation of the recognition result; The label calibration and matching module is used to retrieve a preset existing facility database based on the identification result, the address information, the non-address information, and the vector representation of the identification result, so as to perform comprehensive similarity matching and obtain the similarity matching result of the identification result; The label calibration result acquisition module is used to obtain the calibration result of the recognition result based on the similarity matching result; The existing infrastructure database includes data tables for several types of communication infrastructure, and each data table contains several tag texts and their corresponding vector representations for the same type of communication infrastructure.

10. A computer-readable storage medium, characterized in that, The device includes a stored computer program that, when running, controls the device containing the computer-readable storage medium to perform a tag identification result calibration method for a communication infrastructure as described in any one of claims 1 to 8.

11. A computer program product, characterized in that, It includes a computer program that, when executed by a processor, implements a tag identification result calibration method for a communication infrastructure as described in any one of claims 1 to 8.

12. A terminal device, characterized in that, The system includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor, when executing the computer program, implements a tag identification result calibration method for a communication infrastructure as described in any one of claims 1 to 8.

Images